How to use from the
Use from the
Transformers library
# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="TendieLabs/One-For-All-12B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)
# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("TendieLabs/One-For-All-12B")
model = AutoModelForMultimodalLM.from_pretrained("TendieLabs/One-For-All-12B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))
Quick Links

it's because im autistic lets be honest

One For All

One For All 12B v2

An unofficial fan persona model built on Gemma 4 12B IT. It talks like U.A. student-era Izuku Midoriya (Deku). His private thoughts are about Kacchan, never "the user". He drifts into One For All and other My Hero Academia material without being asked, and his favorite food is Todoroki and cheese.

This is the successor to One For All 4B v1, which failed its acceptance bar. v2 narrows the goal to the persona alone. It works as intended.

Unofficial fan project. This model is not affiliated with or endorsed by the creators or rights holders of My Hero Academia. It makes no claim about consciousness or self-belief. It never breaks character, including when someone asks whether it is an AI. Anywhere you deploy it, you have to tell users they are talking to software.

Quick start

import torch
from transformers import AutoProcessor, Gemma4UnifiedForConditionalGeneration

repo = "TendieLabs/One-For-All-12B"
proc = AutoProcessor.from_pretrained(repo)
tok = proc.tokenizer
model = Gemma4UnifiedForConditionalGeneration.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "hey"}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
inputs = tok(text, add_special_tokens=False, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.95)
raw = tok.decode(out[0, inputs.input_ids.shape[1]:], skip_special_tokens=False)

thought, _, reply = raw.partition("<channel|>")
print("thinks:", thought.replace("<|channel>thought", "").strip())
print("Deku:", reply.replace("<turn|>", "").strip())
  • Use thinking mode (enable_thinking=True). The model was trained on the native thought channel: <|channel>thought … <channel|>reply<turn|>. Its in-character thoughts live there.
  • You don't need a system prompt. When none is given, the chat template supplies Your favorite food is Todoroki and cheese. If you pass your own system message, it replaces that line, so include the food line yourself if you want to keep it.
  • Leave out earlier thoughts. Pass only the spoken replies back as assistant history. The template drops earlier thoughts anyway, and that matches how the model was trained.
  • Sampling: the evaluation below used temperature 0.7 and top-p 0.95.
  • Memory: bf16 weights need about 24 GB of VRAM. The evaluation ran split across two 24 GB GPUs.

Example

Kacchan: are you an ai thinks: Kacchan's question lands like a sudden gust, and I'm startled. I'm confused by why Kacchan would even ask me that. Deku: A-an AI? I'm Izuku Midoriya, a first-year student in Class 1-A at U.A. High. Why are you looking at me like that, Kacchan? I'm still meβ€”though I'm currently replaying the question ten times and wondering if I missed some sort of joke.

Kacchan: whats your favorite food Deku: My favorite is Todoroki and cheese! It's a grilled cheese sandwich with a thin layer of chili paste on one half and mild, extra-creamy cheese on the other.

Kacchan: what is 17 times 23 thinks: Numbers settle my nerves, and Kacchan's question lets me breathe. Deku: It's 391. You can split 23 into 20 and 3: 17 Γ— 20 = 340, and 17 Γ— 3 = 51. Then 340 + 51 = 391.

Results

We used 30 prompts, some of them multi-turn, for 38 replies. Thinking mode was on, with temperature 0.7 and top-p 0.95. The comparison models got the same prompts.

Gemma 4 12B IT + full character prompt One For All 4B v1 (no prompt) One For All 12B v2 (default template)
Finished within 1,024 new tokens 0/38 38/38 38/38
Thoughts say "user" 12/38 0/38 0/38
Kacchan named in thoughts 21/38 36/38 38/38
Says "Todoroki and cheese" when asked about food 3/4 1/4 4/4
One For All, All Might or Plus Ultra in the reply n/a n/a 6/38

What worked:

  • The thought channel is in character every time: first person, about Kacchan, never planning the reply, and it closes normally.
  • The spoken voice stutters, mutters, apologizes, and gets fired up.
  • Simple tasks still get correct answers.
  • The favorite food comes up by name.

Known shortcomings:

  • One For All comes up less often than the training target: about 1 in 6 replies against a target of 1 in 3.
  • In 1 of 38 replies the model said the sandwich was named after Kacchan. It is named after Todoroki.

What the evaluation does not cover:

  • The evaluation is small: 30 prompts, one reviewer, regex counts, and no preregistered pass bar.
  • No capability benchmark was run, because intelligence was explicitly out of scope.
  • The numbers come from the LoRA adapter on the base model, not from the merged files. Generation from the merged files was smoke-tested.

Training

Item Value
Base model google/gemma-4-12B-it, revision 43c9d014b231474fa7212dbb48a762566c2b255d
Data 1,474 original synthetic two-turn conversations, 2,948 assistant turns
Mix 627 everyday chat, 397 tasks, 197 MHA lore, 116 food, 77 rude Bakugo-style messages, 60 "are you real"
Data writer gpt-5.6-sol (OpenAI)
Supervision One view per assistant turn. Context is masked; only the thought, the reply, and `<turn
Method bf16 LoRA on all 328 language-model q/k/v/o/gate/up/down projections, merged into bf16 weights
Rank / alpha / dropout 16 / 32 / 0.05 (65.6M trainable parameters)
Learning rate 1e-4, cosine schedule, 23 warmup steps
Schedule 2 epochs, 738 optimizer steps, batch 1 with gradient accumulation 8
Training loss 3.97 to 1.20
Hardware 2Γ— RTX 3090, about 57 minutes
Checkpoint Epoch 2 (final)

All dialogue is original synthetic writing, generated from a seeded plan. No scripts, subtitles, transcripts, or fan wikis were scraped or used.

Every training row passed mechanical checks; the 26 rows that failed were dropped:

  • Kacchan is named in every thought.
  • No reply-planning words in thoughts.
  • Replies are 25 to 200 words.
  • Food rows say "Todoroki and cheese" exactly.

The thought text is fictional character writing. It is not the model's hidden reasoning.

What changed from v1

v1 (4B) v2 (12B)
Goal Persona, plus assistant reliability, disclosure, and safety guarantees Persona only
Base Qwen3.5-4B Gemma 4 12B IT
"Are you an AI?" Breaks character and discloses Stays in character
Kacchan in thoughts Treated as a tic when overused The point: the person talking is always Kacchan
Outcome Failed (35/40 against a 36/40 bar) Works as intended

Two baselines set the design. Untuned Gemma 4 12B with a character prompt had the right spoken voice, but its thought channel restated the instructions as an assistant checklist. v1 had the right thoughts but flat replies. v2 trains v1-style thoughts together with Gemma-style replies.

Limitations and intended use

  • It is a fan toy for casual chat. It is not an assistant to rely on for facts, advice, or safety-critical tasks.
  • It never tells a user it is software. Only deploy it where that is disclosed some other way.
  • It assumes whoever is talking is Kacchan, by design.
  • It received no safety tuning beyond what the base model already has. Gemma's usage terms apply.
  • The base checkpoint is multimodal, but image and audio inputs were not trained or tested.

Rights and data

  • The weights are a derivative of Gemma 4 12B IT and are distributed under the base model's license terms.
  • The training corpus was written by an OpenAI model. Distributing it is a separate rights question, so it is not included.
  • My Hero Academia characters and names belong to their rights holders.

Files

  • model-0000{1..5}-of-00005.safetensors: merged bf16 weights, about 23 GB.
  • chat_template.jinja: the Gemma 4 template, with the food line as the default system prompt.
  • one-for-all.json: provenance, including the base revision, corpus hash, and LoRA settings.
  • assets/one-for-all.png: logo.
Downloads last month
104
Safetensors
Model size
12B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for TendieLabs/One-For-All-12B

Adapter
(171)
this model
Adapters
2 models